@@ -1,7 +1,19 @@
-# Nightly real-document regression matrix (docs/superpowers/plans/2026-08-15-v9-test-coverage-strategy.md,
-# section 2 tier 2 / section 6). Pulls a public corpus of real Office files
-# (never committed here), drives the real editor through open -> edit -> save
-# for each, and publishes the findings table. Red is a signal, not a gate.
+# What the pull-request gate does not cover, run on a clock instead of on a
+# diff: the PR-time specs on the two engines the gate never launches
+# (playwright.browsers.config.ts), the interaction sweeps and throttled-network
+# budgets, and the real-document matrix over a public corpus of real Office
+# files (docs/superpowers/plans/2026-08-15-v9-test-coverage-strategy.md, section
+# 2 tier 2 / section 6). Red here is a signal, not a gate.
+#
+# Two cadences, one nightly trigger (see the `gate` job):
+#   - cross-browser + sweeps run nightly. They are this repository's ONLY
+#     WebKit/Firefox coverage, and on 2026-08-25 that is what surfaced a
+#     shipped Safari defect the Chromium gate cannot see.
+#   - the corpus matrix runs on Sundays. It is 80 minutes against a corpus that
+#     never changes; what varies is this repository, and a week of it at a time
+#     is enough to keep the findings table current.
+# Neither runs when nothing has been committed since the last time it did --
+# `schedule` fires whether or not there is anything new to say.
 name: Nightly corpus
 
 on:
@@ -20,10 +32,63 @@ on:
         default: "1"
 
 jobs:
+  # Which of tonight's jobs have anything to say. Cheap (one checkout, no
+  # toolchain) and it runs every night, so the run itself is the record of why
+  # a quiet night was quiet -- a workflow that simply stops appearing is
+  # indistinguishable from one that broke.
+  gate:
+    name: What to run tonight
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    outputs:
+      browsers: ${{ steps.decide.outputs.browsers }}
+      corpus: ${{ steps.decide.outputs.corpus }}
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v7
+
+      - name: Decide
+        id: decide
+        env:
+          MANUAL: ${{ github.event_name == 'workflow_dispatch' }}
+        run: |
+          age_hours=$(( ( $(date -u +%s) - $(git log -1 --format=%ct) ) / 3600 ))
+          # 1..7 with 7 = Sunday. The cron fires at 19:17 UTC, so this is the
+          # UTC weekday of the trigger, not of the Asia/Shanghai morning it
+          # lands in -- which is what the comment above the crons means.
+          weekday=$(date -u +%u)
+
+          browsers=false
+          corpus=false
+          # A cycle plus an hour of slack on each window: a commit landing just
+          # before the run it belongs to would otherwise wait a whole cycle.
+          # Written as `if` blocks rather than `a || b && c` -- that list
+          # evaluates to non-zero when both tests fail, and the step runs under
+          # `bash -e`, so a quiet night would end as a failed gate.
+          if [ "$MANUAL" = "true" ] || [ "$age_hours" -lt 25 ]; then
+            browsers=true
+          fi
+          if [ "$MANUAL" = "true" ] || { [ "$weekday" -eq 7 ] && [ "$age_hours" -lt 193 ]; }; then
+            corpus=true
+          fi
+
+          {
+            echo "HEAD is ${age_hours}h old, UTC weekday ${weekday}, manual=${MANUAL}"
+            echo ""
+            echo "- cross-browser + sweeps: ${browsers}"
+            echo "- corpus matrix: ${corpus}"
+          } >> "$GITHUB_STEP_SUMMARY"
+
+          echo "browsers=${browsers}" >> "$GITHUB_OUTPUT"
+          echo "corpus=${corpus}" >> "$GITHUB_OUTPUT"
+
   corpus:
     name: Real-document matrix (Apache POI test-data)
     runs-on: ubuntu-latest
     timeout-minutes: 300
+    needs: gate
+    if: needs.gate.outputs.corpus == 'true'
 
     steps:
       - name: Checkout code
@@ -83,6 +148,8 @@ jobs:
     name: Cross-browser (WebKit + Firefox)
     runs-on: ubuntu-latest
     timeout-minutes: 90
+    needs: gate
+    if: needs.gate.outputs.browsers == 'true'
 
     steps:
       - name: Checkout code
@@ -96,7 +163,17 @@ jobs:
       # The same PR-time suites (minus the opt-in sweeps) on the two engines
       # the gate does not cover. Same-realm/L0 rules apply unchanged.
       - name: Run E2E on WebKit and Firefox
-        run: pnpm exec playwright test -c playwright.browsers.config.ts --grep-invert "api surface|corpus" --reporter=list
+        run: pnpm exec playwright test -c playwright.browsers.config.ts --grep-invert "api surface|corpus|@serial" --reporter=list
+
+      # The other half of that `--grep-invert`, exactly as ci.yml's e2e job
+      # pairs them: an @serial case is one that cannot share the run. Without
+      # this split sw-silent-update.spec.ts rewrites the served sw.js while
+      # other specs are being served from it, replacing THEIR worker mid-test
+      # -- which is where sw-warm's "runtime-e2e-next" and cache-first's
+      # missing sentinel came from on 2026-08-25. Without the second pass
+      # nothing runs them at all and the job stays green regardless.
+      - name: Run the cases that cannot share the run
+        run: pnpm exec playwright test -c playwright.browsers.config.ts --grep @serial --workers=1 --reporter=list
 
       - name: Upload report
         if: always()
@@ -110,6 +187,8 @@ jobs:
     name: Slow-network budgets
     runs-on: ubuntu-latest
     timeout-minutes: 30
+    needs: gate
+    if: needs.gate.outputs.browsers == 'true'
 
     steps:
       - name: Checkout code
 
@@ -224,9 +224,10 @@ test/setup/vitest.ts          # 全局 mock：matchMedia、URL.createObjectURL
   `CORPUS_LIMIT=<上限，截断会打印>` / `CORPUS_VISUAL=1`（L3：原始 vs 存回再
   打开的渲染逐像素比对）；每条用例结束即追加一行到
   `test-results/corpus-rows-<worker>.jsonl`，`node bin/corpus-report.mjs [dir]`
-  合并出 JSON + markdown 表（含 L2 内容、L3 视觉、open/save p50/p95）；**夜间 CI** `.github/workflows/nightly-corpus.yml`
+  合并出 JSON + markdown 表（含 L2 内容、L3 视觉、open/save p50/p95）；**CI 每周日**
+  由 `.github/workflows/nightly-corpus.yml`
   拉 Apache POI test-data 公开语料跑同一套（红=信号非门禁，可
-  workflow_dispatch 指定 limit/filter）；对目录下每个 docx/doc/xlsx/xls/pptx/ppt/csv 走
+  workflow_dispatch 随时手动触发并指定 limit/filter）；对目录下每个 docx/doc/xlsx/xls/pptx/ppt/csv 走
   打开 → 监听致命弹窗/asc_onError → 编辑 → 保存 → 产物 sanity，输出汇总
   报告。语料留在测试机上不入库；未设 `CORPUS_DIR` 整套 skip，CI 保持绿。
   第一天就抓出 P0（非 ASCII 文件名导致 -82 打开失败 + 永久转圈），见
@@ -251,6 +252,19 @@ test/setup/vitest.ts          # 全局 mock：matchMedia、URL.createObjectURL
 （WebKit + Firefox，夜间）、`playwright.prod.config.ts`（打线上/preview，
 `prod-smoke.yml` 与 PR 门禁 `preview-smoke.yml` 用）。
 
+**夜间那个 workflow 有两档节奏，且没有新提交就整晚不跑**（`nightly-corpus.yml`
+的 `gate` job，2026-08-25 起）：cross-browser + sweeps **每天**（它是本仓库唯一
+的 WebKit/Firefox 覆盖，PR 门禁全是 Chromium），corpus 矩阵**每周日**（80 分钟，
+而语料是静态的，变的是本仓库）；`workflow_dispatch` 无视这两个判断。判据是
+HEAD 的提交时间，窗口各留一个周期加一小时的余量。**gate 本身每晚都跑**——安静
+的夜晚也要留下一条 run 说明为什么安静，否则"没有 run"和"workflow 坏了"长得一样。
+**cross-browser 那个 job 必须和 ci.yml 一样把 `@serial` 分成两趟**
+（`--grep-invert "api surface|corpus|@serial"` + `--grep @serial --workers=1`）：
+`sw-silent-update.spec.ts` 靠改写被服务的 `dist/sw.js` 来模拟一次部署，那是**整个
+源站可见**的，跟别的 spec 并排跑会把它们的 worker 也换掉——2026-08-25 之前
+`sw-warm` 读到 `runtime-e2e-next`、`sw-vendor-cache-first` 找不到自己刚种下的
+哨兵，都是这个。两半必须成对，`workflow-contract.test.ts` 钉住。
+
 **L0 全局 fixture（`test/e2e/lib/l0.ts`，2026-08-15 起所有 spec 从它
 导入 `test`/`expect`）**：自动把 `asc_onError`、厂商致命弹窗、编辑器 iframe
 内的 `unhandledrejection`/`error`、pageerror、非白名单 console.error 判为
 
@@ -377,6 +377,35 @@ describe('.github/workflows/nightly-corpus.yml', () => {
     expect(readFileSync(resolve(ROOT, 'test/e2e/lib/l0.ts'), 'utf8')).toMatch(/projectName: \[/);
   });
 
+  it('keeps the cases that cannot share the run out of the parallel pass, and still runs them', () => {
+    // The same pairing ci.yml's e2e job has, for the same reason and with the
+    // same failure mode when either half goes missing. Here the case that
+    // needs it is sw-silent-update.spec.ts: it rewrites the served sw.js to
+    // stand in for a deploy, which replaces the worker under every other spec
+    // on the origin -- sw-warm read a vendor stamp that belonged to it, and
+    // the cache-first probe lost the entry it had just planted.
+    expect(src).toMatch(/--grep-invert "api surface\|corpus\|@serial"/);
+    expect(src).toMatch(/-c playwright\.browsers\.config\.ts --grep @serial --workers=1/);
+  });
+
+  it('runs each half of the night only when it has something to say', () => {
+    // `schedule` fires whether or not anything was committed, and the corpus
+    // is static: what varies is this repository. Both windows are decided in
+    // one cheap gate job so a quiet night still leaves a run explaining itself.
+    const jobs = parseJobs(src);
+    expect(jobs.map(({ name }) => name)).toContain('gate');
+    for (const name of ['browsers', 'budgets']) {
+      const job = jobs.find((candidate) => candidate.name === name);
+      expect(job, name).toBeDefined();
+      expect(job!.body, name).toMatch(/^ {4}if: needs\.gate\.outputs\.browsers == 'true'$/m);
+      expect(job!.body, name).toMatch(/^ {4}needs: gate$/m);
+    }
+    const corpus = jobs.find(({ name }) => name === 'corpus');
+    expect(corpus!.body).toMatch(/^ {4}if: needs\.gate\.outputs\.corpus == 'true'$/m);
+    // Manual runs always go through, whatever the clock says.
+    expect(src).toMatch(/MANUAL: \$\{\{ github\.event_name == 'workflow_dispatch' \}\}/);
+  });
+
   it('drops the corpus fuzzer output without dropping real bug-report files', () => {
     // Byte soup minimized by a fuzzer until it broke a parser: "this editor
     // will not open it either" is not a finding, and 15 such files were red on